Single-cell RNA sequencing via endoscopic ultrasound-guided fine-needle biopsy for pancreatic tumors uncovered an aggressive subclone with a poor prognosis

Abstract

Background

Single-cell RNA sequencing (scRNA-seq) is a powerful tool which can unveil regulatory connections between genes and cells. However, due to the high demand for tissue quality, most scRNA-seq for pancreatic cancer is carried out by surgical specimen. This study (NCT05767697) aims to gain practical experience in carrying out scRNA-seq using endoscopic ultrasound-guided fine-needle biopsy (EUS-FNB).

Materials and methods

With a within-subjects design, two punctures from the same lesion, with a negative pressure of 5 ml (suction group) and without applying suction (non-suction group), were evaluated. Each biopsy sample was separated into three parts: white tissue, red material, and buffer for living cell counting. scRNA-seq was carried out according to the manufacturer’s protocol. The pancreatic adenocarcinoma dataset in The Cancer Genome Atlas (TCGA) was used as external validation.

Results

A total of 20 patients with 40 specimens were enrolled. The suction group achieved a success rate of 80% (16/20), which was significantly higher than the 10% (2/20) success rate in the non-suction group (P < 0.001). scRNA-seq data were generated for four patients, including two early stage and two late stage. Overall, 15 major cell subtypes, including 4 cancer subclones, were identified across early and late stages. Analysis of differentially expressed genes identified an aggressive subclone highlighted by ubiquitin-conjugating enzyme E2 C (UBE2C) high expression, commonly in late stage. The TCGA dataset also demonstrated that UBE2C high-expression pancreatic cancer had a poor prognosis.

Conclusions

EUS-FNB with a negative pressure of 5 ml is feasible for scRNA-seq in clinical practice. A UBE2C high-expression subclone correlates with a poor prognosis, potentially becoming a new therapeutic target in future studies.

Highlights

• •EUS-FNB is feasible for scRNA-seq in clinical practice and will serve as a powerful tool for pancreatic cancer research.
• •A UBE2C high-expression subclone also exists in early-stage pancreatic cancer.
• •UBE2C high expression correlates with a poor prognosis.

Introduction

Since its introduction in 1992, endoscopic ultrasound-guided fine-needle aspiration/fine-needle biopsy (EUS-FNA/FNB) has become the standard for diagnosing solid pancreatic masses.¹ The diagnostic sensitivity of EUS-FNA/FNB for pancreatic cancer increases as the operator’s experience level rises after 30 procedures.² Meta-analyses revealed an 85%-89% sensitivity and a 96%-99% specificity for the diagnosis of solid pancreatic mass.^3,4 Nowadays EUS-FNA/FNB offers comparable diagnostic performance to other methods (e.g. percutaneous biopsy⁵) with a very low risk of needle tract seeding (∼0.1%)⁶ and EUS-FNB before surgery did not affect cancer-specific survival.⁷ It is important as neoadjuvant therapy is recognized as the standard of care in borderline resectable pancreatic cancer^8,9 and locally advanced pancreatic cancer10, 11, 12, 13 with increasing evidence of benefit in resectable pancreatic cancer.¹⁴

In the era of precision medicine, EUS-FNB is utilized not only for diagnosis but also for genetic testing. The mutational landscape of pancreatic ductal adenocarcinoma (PDAC) has been reported in prior research, identifying KRAS, TP53, CDKN2A, and SMAD4 as the most common genetic alterations in PDAC. However, due to the high demand for adequate tissue specimens for comprehensive genomic profiling, this genetic information generally comes from surgical specimens.^15,16 This raises the question: can specimens from resectable PDAC represent all PDAC, given that resectable PDAC only accounts for ∼20% of all PDAC? Furthermore, the gene expression measured by conventional bulk RNA sequencing, as used in The Cancer Genome Atlas (TCGA) project,¹⁷ can only provide an average expression signal for a group of cells, not individual cells, and therefore might introduce ecological bias.¹⁸ Single-cell RNA sequencing (scRNA-seq) can unveil intricate and uncommon cell populations, unveil regulatory connections between genes, and is rapidly involved in novel cancer research.¹⁹ However, due to the high demand for tissue quality and cell viability, currently most scRNA-seq is carried out by surgical specimen or biopsy from metastatic sites.20, 21, 22, 23

In clinical practice, the acquired substance from EUS-FNB is predominantly composed of red material rather than white tissue. In our previous studies, utilizing digital pathology assistance and next-generation sequencing (NGS), we demonstrated that red material had a comparable tumor area with white tissue, suggesting that both white tissue and red material were feasible for genomic profiling.^24,25 This study aims to establish a practical experience to help scientists carry out scRNA-seq using EUS-FNB samples in daily clinical practice.

Materials and methods

Study design

This is a prospective comparison study with a within-subjects design that took place in a high-volume center for pancreatic cancer in southern Taiwan. The inclusion criteria were as follows: (i) solid pancreatic lesions with suspicious malignancy by cross-sectional imaging; (ii) confirmation of pancreatic cancer by cytology/histology was essential for the selection of appropriate treatment; (iii) willingness to provide fresh tissue for research. The exclusion criteria included (i) those unable to cooperate with the procedure; (ii) coagulopathy defined as a platelet count <50 000/ml or a prothrombin time prolonged for >3 s. This study is approved by the investigational review board of National Cheng Kung University Hospital (A-ER-109-073) and is registered in ClinicalTrials.gov (NCT05767697). Informed consent was obtained from all patients involved in the study.

Sample size justification

The sample size for this study was not determined through statistical analysis. Instead, it was set at 20 cases (10 per group), based on historical experience with safety and tolerance, while also considering the feasibility of completing the study within the designated timeframe.

EUS-FNB

In contrast to percutaneous core biopsies, which involve cutting a tissue cylinder, EUS-FNB sampling utilizes a syringe that applies negative pressure and forward motion to aspirate cells. However, this technique carries the risk of significant blood contamination which would be detrimental to the subsequent scRNA-seq. To this end, we compared two protocols to obtain EUS-FNB material with minimal blood contamination, which is promptly followed by an assessment of sample quality.

All procedures were carried out by a highly experienced physician, M-YL, who had conducted over 500 EUS-FNA/FNB procedures. The patients underwent EUS using a linear echoendoscope (GF-UCT260; Olympus, Tokyo, Japan), which was connected to either the EU-ME1 or EU-ME2 system (Olympus). The EUS procedure utilized a 22-gauge Acquire needle (Boston Scientific, Marlborough, MA). Depending on the location of the target tumor, the needle was carefully inserted into the pancreatic mass through either the stomach or the duodenum, ensuring the absence of nearby blood vessels. Each needle pass involved the fanning technique, employing 16-20 door-knocking strokes under direct visualization of the needle tip using EUS. Each participant typically underwent 2-4 punctures under a negative pressure of 10 ml for routine histopathological examination and clinical diagnosis. Following this, an additional two punctures from the same lesion were taken using the same needle—one with a negative pressure of 5 ml (suction group) and one without suction (non-suction group). These samples were directly transported to the laboratory for cell preparation without histopathological evaluation.

Cell preparation

The tissue acquired from EUS biopsies was placed in a 10-cm dish on ice and transferred to the laboratory within 1 h. Samples from each condition (without suction or with a negative pressure of 5 ml) were separated into three parts, white tissue, red material, and buffer (Figure 1A and B). Each part then proceeded to a different subsequent cell preparation protocol. Briefly, cells were incubated with enzyme digestion buffer (Miltenyi Biotec, Bergisch Gladbach, Germany) separately for 15 min to generate single-cell suspension according to the manufacturer’s protocol. Single-cell suspension was centrifuged at 3000 rpm (856 g) for 1 min to collect cell pellets and incubated with red blood cell lysis buffer for 2 min at room temperature to remove erythrocytes. Cells were then centrifuged at 1000 rpm (96 g) for 3 min to collect cell pellets and resuspended in Dulbecco’s modified Eagle’s medium (Hyclone™, Cytiva, Marlborough, MA) supplemented with 10% fetal bovine serum (Hyclone™). Cell viability was stained by trypan blue staining (Invitrogen, Waltham, MA), and measured with TC20™ automated cell counter (Bio-Rad, Hercules, CA) to ensure that the cell viability was >70%. Samples with cell viability <70% were to undergo dead cell removal to increase cell viability.

Figure 1.

Sample condition. (A) Suction group. (B) Non-suction group. (C) Total retrieved live cells in the suction group and non-suction group. Dots above the red dashed line represent an adequate number of live cells for subsequent single-cell RNA sequencing. Blue dots represent trans-duodenal biopsy, while red dots represent the trans-gastric approach.

A total of 10 000 cells with a cell viability of ≥70% are recommended by the manufacturer 10X and samples with a cell viability of <70% may undergo dead cell removal. Given that the dead cell removal step may slightly increase cell viability but may result in a significant loss of cells,²⁶ we defined the success of cell preparation as a total live cell number of ≥100 000 and a cell viability of ≥60%.

Library preparation and sequencing

scRNA-seq libraries were prepared with Chromium Single Cell 5′ Reagent Kits v2 (10X Genomics, Pleasanton, CA) following the manufacturer’s protocol. In brief, 10 000 cells which were mixed with reverse transcription reagent, gel beads, and partitioning oil were loaded into a Chromium Next GEM Chip K. Then, Chip K was subsequently loaded into a Chromium Controller to generate gel bead-in-emulsions and followed by reverse transcripts (53°C for 45 min; 85°C for 5 min) to produce complementary DNA (cDNA). cDNA was then cleaned up with Dynabeads and followed by cDNA amplification [98°C for 45 s (98°C for 20 s, 63°C for 30 s, and 72°C for 1 min) × 11 cycles; 72°C for 1 min]. The amplified cDNA was then fragmented, ligated with adapter and sample index, and double-sided selection carried out with solid-phase reversible immobilization beads (Beckman, Brea, CA) according to the manufacturer’s protocol. Finally, the libraries were quantified with the Qubit dsDNA HS assay kit (ThermoFisher, Waltham, MA), and the library’s average fragment size was measured with the 4150 TapeStation system (Agilent, Santa Clara, CA). Finally, libraries were sequenced on an Illumina NextSeq 2000 platform (Illumina, San Diego, CA). A total of 2 × 150 bp paired-end runs of Illumina NextSeq 2000 were carried out according to the manufacturer’s protocol.

Sequencing data processing and analysis

Sequencing data were imported to Cellranger (version 6.0.1, 10x Genomics, Pleasanton, CA) separately for further analysis. We used the ‘cellranger mkfastq’ function to demultiplex base-call files into FASTQ files, followed by the ‘cellranger count’ function to carry out alignment, filtering, barcode counting, and unique molecular identifier counting. We used the Seurat package (version 4.0.3) in the R program for further analysis. To remove possible low-quality and doublet cells, we excluded the cells with <200 genes and >4000 genes in our dataset; the percentage of mitochondrial genes which were >20% was also excluded due to potential apoptotic cells. We integrated our samples with the standard integration protocol as described in the Seurat package to remove the batch effect correction between each dataset. Individual datasets were normalized with the ‘NormalizeData’ function and 2000 highly variable genes were selected with the ‘FindvariableFeatures’ function under default parameters. We then used the ‘FindIntegrationAnchors’ function to identify the integration anchors between each dataset followed by the ‘IntegrateData’ function to integrate all datasets. Principal component analysis (PCA) was carried out with the ‘RunPCA’ function and the amount of included PCA components (PCs) was evaluated with elbow plot. Within all the PCs, the top 25 PCs were used for clustering and Uniform Manifold Approximation and Projection (UMAP) analysis. Differentially expressed genes in each cluster were identified with log2-fold change >0.25 in the Seurat ‘FindAllMarkers’ function. Cell type identification of each cluster was determined by the SingleR package (version 1.7.1) and manually checked with known cell surface markers from the Panglaodb database (https://panglaodb.se/index.html) and literature review. The marker genes to identify cell populations in our study are shown in Supplementary Table S1, available at https://doi.org/10.1016/j.esmogo.2024.100113.

External validation and statistical analysis

To further substantiate and validate our findings, we conducted an external validation using comprehensive datasets from the TCGA database. The data, generated using Illumina HiSeq technology, were normalized using the RNA-Seq by Expectation Maximization method for gene expression quantification. Specifically, we obtained the pancreatic adenocarcinoma (PAAD) RNA sequencing data and corresponding clinical information, including patient demographics and survival outcomes, from OncoLnc (http://www.oncolnc.org/).

The Kolmogorov–Smirnov test was employed to assess the normality of each variable. Fisher’s exact test was utilized for comparing count data between groups. To compare continuous variables between groups, the Wilcoxon rank sum test and the Kruskal–Wallis test were applied. Variables with a P value <0.05 were deemed statistically significant. All statistical analyses were conducted using R version 4.1 (R Core Team, Vienna, Austria).

Results

Patient characteristics

A total of 20 patients were enrolled. The median age was 65.1 years (range 46.6-83.2 years), with an equal distribution between males and females (Table 1). The majority (60%) of lesions were located at the pancreatic head and the median lesion size was 2.6 cm (range 0.9-6.7 cm). Eighteen of 20 patients (90%) were diagnosed with advanced stage including 7 with stage III and 11 with stage IV (the American Joint Committee on Cancer, AJCC, 8th edition). The trans-gastric and trans-duodenal needle routes were used in each of 10 patients, respectively. Nineteen were confirmed to have adenocarcinoma, while one patient was diagnosed with carcinoma in situ. A definitive diagnosis of adenocarcinoma for the latter patient could not be made due to fragmented tissue lacking sufficient stromal components to assess for stromal invasion (Table 1).

Table 1.

Patient characteristics

	Overall (N = 20)
Age, median (range), years	65.1 (46.6-83.2)
Sex, n (%)
Female	10 (50.0)
Male	10 (50.0)
Lesion location, n (%)
Head	12 (60.0)
Body	5 (25.0)
Tail	3 (15.0)
Lesion size, cm, median (range)	2.6 (0.9-6.7)
Stage, n (%)
I	1 (5.0)
II	1 (5.0)
III	7 (35.0)
IV	11 (55.0)
Final diagnosis, n (%)
Adenocarcinoma	19 (95.0)
Carcinoma in situa	1 (5.0)
Histology grade, n (%)
Well differentiated	2 (10)
Moderately differentiated	16 (80)
Poorly differentiated	2 (10)
Needle route, n (%)
Trans-gastric	10 (50.0)
Trans-duodenum	10 (50.0)

^aA definitive diagnosis of adenocarcinoma could not be made due to fragmented tissue lacking sufficient stromal components to assess for stromal invasion.

Cell number and viability

As expected, suction with a negative pressure of 5 ml (Figure 1A) had more blood contamination as compared with those without suction (Figure 1B). As a trade-off, negative suction yielded significantly more tissue than those without suction. Recovery of 100 000 live cells was the qualification criterion for scRNA-seq. In the suction group, 12 out of 20 (60%) buffer samples and 11 out of 20 (55%) red material samples were qualified, with median total live cell numbers of 3.4 × 10⁵ cells and 1.8 × 10⁵ cells, respectively, while only 1 out of 20 (5%) white tissue samples was qualified, with a median cell number of 7.9 × 10³ cells. In contrast, only 2 out of 20 (10%) buffer samples qualified in the no suction group, with a median cell number of 1.3 × 10⁴ cells, and no samples in either the red material or white tissue groups were qualified, with median cell numbers of 7.0 × 10³ cells and 3.9 × 10³ cells, respectively (Figure 1C). There was no significant difference in live cell recovery between the trans-duodenal and trans-gastric approaches (Figure 1C). Overall, combining all parts of samples, suction with a negative pressure of 5 ml achieved a cell preparation success rate of 80% (16/20) which was significantly higher than the 10% (2/20) success rate in the non-suction group (P < 0.001). The total live cell counts were not significantly different among histology grade, with medians of 6 × 10⁴, 1.0 × 10⁵, and 1.3 × 10⁵ cells in the well-differentiated, moderately differentiated, and poorly differentiated categories, respectively, in the suction group (P = 0.44). In contrast, the corresponding non-suction group showed medians of 1.3 × 10⁴, 7.0 × 10³, and 5.7 × 10³ cells, respectively (P = 0.96) (Supplementary Figure S1, available at https://doi.org/10.1016/j.esmogo.2024.100113).

Sample quality assessment

After establishing the cell preparation protocol for EUS-FNB, we utilized the 10X Genomics Chromium platform to generate single-cell RNA transcriptomes for four patients. A total of 14 224 cells, including two early-stage samples (9632 cells) and two late-stage samples (4592 cells) were extracted. After excluding potentially low-quality and empty droplet cells, a total of 11 950 cells were utilized for downstream analysis. In total, 66% of all single cells (7842) belonged to the early-stage dataset, while the remaining 34% (4108) belonged to the late-stage dataset.

Seurat unsupervised cell cluster analysis

We employed the Seurat v4.0.3 R package for further analysis. After integrating all datasets, we conducted PCA using the top 2000 highly variable genes. The first 25 principal components were selected as input for the nonlinear dimensionality reduction algorithm UMAP to visualize cell clusters through the elbow plot (Supplementary Figure S2, available at https://doi.org/10.1016/j.esmogo.2024.100113).

Our analysis confirmed that the integration method successfully removed batch effects and demonstrated a high degree of reproducibility across each dataset (Supplementary Figure S3, available at https://doi.org/10.1016/j.esmogo.2024.100113). Using a modularity optimization technique, cells were clustered, and the optimal resolution parameter of 0.5 was determined through optimization within a range from 0.3 to 1.1, leading to the identification of a total of 19 clusters in our single-cell dataset (Supplementary Figure S4, available at https://doi.org/10.1016/j.esmogo.2024.100113).

For cell type annotation, we utilized the SingleR package with built-in references from the Human Primary Cell Atlas and Blueprint and Encode. Combining this information with differentially expressed gene profiles and insights from existing literature, we successfully identified 15 major cell subtypes across both early- and late-stage datasets, which included CD4+ T cells (IL7R, LTB, CD3D), CD8+ T cells (CD8A, GZMK), regulatory T cells (FOXP3, IL2RA), natural killer cells (GNLY, KLRD1), monocytes (CD14, FCN1), macrophages (APOE, CTSB), neutrophils (CMTM2, FCGR3B), B cells (MS4A1, CD79A), plasma cells (JCHAIN, CD79A), mast cells (MS4A2, GATA2), fibroblasts (COL1A1, THY1), and four different subtypes of cancer cells (KRT8, KRT18), respectively (Figure 2A and B).

Figure 2.

Tumor heterogeneity. (A) The UMAP of the merged embedding of the four scRNA-seq samples shows four clusters of cancer cells. (B) The proportion of each cell types in two early-stage and two late-stage patients.

NK, natural killer; scRNA-seq, single-cell RNA sequencing; Treg, regulatory T; UMAP, Uniform Manifold Approximation and Projection.

The proportion of cancer cells-1 was significantly higher in the early-stage samples, while cancer cells-4 showed a significant increase in the late-stage samples (Figure 3A). Analysis of differentially expressed genes in all cancer cells revealed that MKI67 and UBE2C were the most highly expressed genes in cancer cells-4, suggesting that this subclone is highly proliferative (Figure 3B and C).

Figure 3.

Transcriptional subtype heterogeneity. (A) The distribution of cancer subclones. (B) Expression of the top five enriched genes for each cluster. (C) Expression of MKI67 and UBE2C for each cluster.

UBE2C, ubiquitin-conjugating enzyme E2 C.

External validation in TCGA

A total of 177 patients diagnosed with PDAC, with comprehensive data on survival outcomes and transcriptome information, were included. The messenger RNA expression level of ubiquitin-conjugating enzyme E2 C (UBE2C) was significantly higher in patients with T3/T4 tumors compared with those with T1/T2 tumors (P = 0.0153; Figure 4A). The median overall survival of UBE2C low-expression tumors was 23.1 months [95% confidence interval (CI) 19.8 months-not estimable], which was significantly better than that of UBE2C high-expression tumors, with a median overall survival of 15.9 months (95% CI 12.9-22.5 months, P = 0.0016; Figure 4B).

Figure 4.

External validation of TCGA database. (A) The expression of UBE2C in patients with T1/T2 and T3/T4 tumors, respectively. (B) The Kaplan–Meier plot of patients with UBE2C low expression and high expression.

CI, confidence interval; NE, not estimable; TCGA, The Cancer Genome Atlas; UBE2C, ubiquitin-conjugating enzyme E2 C.

Discussion

EUS-FNB with a negative pressure of 10-20 ml is commonly used in daily practice for solid pancreatic tumor diagnosis.²⁷ However, our previous experience suggested that a 10-ml negative pressure usually resulted in significant blood contamination so in the present study we applied a negative pressure of 5 ml instead.²⁵ A previous study demonstrated that utilizing two passes with a 22-gauge Franseen needle provided sufficient tissue for the extraction of RNA and DNA, facilitating subsequent NGS in nearly 95% of patients diagnosed with pancreatic cancer.²⁸ Our study not only confirmed that, in daily practice, two passes of EUS-FNB are generally sufficient for conventional bulk DNA and RNA sequencing but also demonstrated the feasibility of utilizing two passes of EUS-FNB for scRNA-seq. The potential application of EUS-FNB with a negative pressure of 5 ml for scRNA-seq in routine clinical practice represents a significant advancement in the field. This technique not only enhances our ability to obtain high-quality samples but also opens up new avenues for exploring the molecular landscape of pancreatic cancer at the single-cell level.

While the major focus of this study is to establish a practical protocol for scRNA-seq using EUS-FNB in daily practice, our scRNA-seq analysis also identifies a subclone characterized by high expression of UBE2C even in the early stages of pancreatic cancer. UBE2C, a gene involved in cell cycle regulation and ubiquitination processes, is scarcely detectable in normal tissues.²⁹ In contrast, a growing body of evidence points toward the overexpression of UBE2C in various human tumor types, such as thyroid cancer,³⁰ breast cancer,³¹ gastric cancer,³² and pancreatic cancer.³³ Notably, a previous study demonstrated that UBE2C was required to induce tumorigenesis for lung cancer with KRAS mutations.³⁴ UBE2C is also associated with tumor progression and epithelial–mesenchymal transition, a key driver of cancer metastasis, in several cancer types including pancreatic cancer.^30,33,35 Our observation of a UBE2C high-expression aggressive subclone in early-stage pancreatic cancer highlights the power of scRNA-seq in studying the complicated tumor heterogeneity and microenvironment of PDAC. The identification of the UBE2C high-expression subclone and its association with a poor prognosis suggests avenues for further research, with the potential for developing targeted therapies.

Our study had some limitations. Firstly, owing to the limited number of cases, we were unable to analyze survival outcomes in our patient cohort. Therefore, the role of the UBE2C high-expression subclone in early-stage pancreatic cancer should be further validated in a larger scRNA-seq cohort with a greater number of cases. Secondly, the detailed mechanism of how UBE2C regulates epithelial–mesenchymal transition and promotes aggressive subclone cancer cell metastasis is not well recognized in the current study. However, since the current study primarily aims to establish practical experience in using EUS-FNB for scRNA-seq, the detailed mechanism study is not the primary focus of the current study and will be conducted in future research.

In conclusion, we have developed a practical protocol demonstrating the feasibility of using a negative pressure of 5 ml for scRNA-seq in clinical settings. Additionally, our analysis of scRNA-seq data identified an aggressive subclone, highlighting the potential of this approach in elucidating tumor heterogeneity.